Agent Designer
A Pi skill (v0.3.0, MIT) that designs and deploys expert ensembles: casts of LLM-generated domain-expert personas coordinated through a fixed, inherited methodology. It canonizes the craft behind eight hand-built persona panels in the author's experiments/ai-simulations — the skill extracts their shared orchestration spine into a versioned kernel and never touches the originals again.
Two protocols share one craft layer. The panel protocol builds a standing panel — moderator, 5–8 specialist seats, one permanent adversarial/audit seat, phase overlays — that lives at panels/<slug>/ in the repo where the work happens, discovered by find . -name panel.json rather than any central registry. A calibration gate flips a manifest from calibrating to active only after the user confirms the goal and roster; nothing is drafted before that. The runtime protocol casts 2–5 experts live for one question, executes them through Pi's subagent tool (parallel isolated passes, then an adversarial pass, then synthesis), and leaves transcripts plus a promotable cast on disk — promotion absorbs the files, it never regenerates them.
The persona doctrine's theory is that a persona is a retrieval key into the training distribution: a domain-common name, an institutional email signature, a CV in the field's own vocabulary, a specific formative experience, declared proclivities and blind spots, textured communication style, and an expert query vocabulary each land the model in expert-text territory, while generic pointers keep it in generic-assistant space. A two-pass gate enforces this: validate-persona.sh checks structure (eight sections, ≥300 words), and a model-mediated voice checklist re-drafts anything that "reads like a job posting" or has "no vocabulary of its own".
The orchestration canon (methodologies/kernel.md, 11 rules, semver-versioned, currently 1.2.0) is inherited whole — overlays add phase work but may never amend it. Its distinctive moves: specialists work in isolation before convergence (anti-anchoring), disagreement is signal and every synthesis must carry a labeled ## Dissent section or a script fails it, quality judgments are model-mediated while deterministic scripts check only structure, and every run closes with a contribution review where causal-lift claims require a one-seat counterfactual — otherwise only traceable contribution may be claimed. A 2026-09-09 self-critique session produced seven foundational challenges to the doctrine; the repo answered with a calibration addendum and an opt-in assignment-decomposition pilot overlay rather than silent edits.